A Privacy-Preserving and Byzantine-Resilient Federated Learning Scheme in VANETs

With the widespread adoption of federated learning in vehicular ad-hoc networks (VANETs), gradient privacy leakage and Byzantine attacks have emerged as critical research challenges. To address these issues, this paper proposes a privacy-preserving and Byzantine-resilient federated learning scheme in VANETs. Specifically, we use obfuscation factors to obfuscate the local gradients to ensure their privacy is not compromised. Then, we propose a method that calculates vehicles’ credibility using local gradient deviation, and this method identifies Byzantine nodes by evaluating the degree of gradient deviation in VANETs. Security analysis and experimental results show that the proposed scheme can protect the privacy of vehicles and also identify Byzantine nodes to improve the accuracy of the global model.